Liquid AI Releases LFM2.5-24B-A2B: A New Reasoning Model

LFM2.5-24B-A2B, a 24B-parameter reasoning model from Liquid AI, is engineered for swift, private agentic tasks on-device, utilizing only 2B active parameters per token. This text-only model supports a context window of 131,072 tokens and is optimized for reasoning, tool use, and agentic workflows, though it is recommended to pair it with retrieval or web search for knowledge-heavy tasks. Developers can deploy LFM2.5-24B-A2B locally using various formats like Transformers, vLLM, SGLang, llama.cpp, and ONNX, or on Apple Silicon with MLX. Liquid AI claims the model is competitive with Qwen3.6-35B-A3B on agentic tasks, achieving a score of 48.09 on LiveClaw, and runs up to 1.7× faster than comparable MoEs on-device, delivering 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU. The model’s versatility is further enhanced by its support for multiple languages and its tool-calling capabilities, making it suitable for a wide range of applications.